GTA6-like as a 942Kb UEFI bootable image (YouTube explainer, Hacker News thread) showed up on my news feed recently. Whether or not it was vibecoded, it’s still a technical feat of compression. For comparison, GHC produces a 1.3MB “Hello, world!” executable, and Rust does the same job in 440KB. The demoscene was on my mind after seeing this. This morning, it collided with some thoughts about software pricing in a fun way.
I like the idea of selling software along with its source code for a flat fee. One of the FAQs I imagine in 2026 is, “Why would I buy this when I could use an LLM to build my own?” My current answer would be, “If that makes economic sense, then go ahead!”
“Build instead of buy” has always been an option. If you know what someone’s program does, even only at a high level, then with enough time and effort you can make one that does the same thing. When the predicted time and effort is high enough, or uncertain enough, buying is a good idea. LLMs significantly reduce the time and effort required to copy a program’s features. They also make it easier to quantify the cost of producing software, because you pay by the token. So one way to check the price of a program is to consider how much it would cost, in tokens, to replicate using an LLM. If I’m selling a program for $10 but it could be reliably replicated by a state-of-the-art LLM for $1, then I’ll be undercut. But if instead it cost $100 in tokens to replicate the program, then buying is still a good deal.
Thinking about this, I realised that I don’t have a good sense of any model’s potential given a fixed budget. Demos often have a program size limit, such as 4KiB or 64KiB, which forces programmers to squeeze their platform for everything it’s got. I’m curious about what would happen if people competed in LLM program extrusion under a particular token budget. For example, what’s the most technically challenging thing you can build with Claude Fable 5.2 for $10 (Nov 2026 pricing)? How about $100? Can you make 10x more with 10x the budget? What would we learn about these systems by trying to push them to their limits, rather than simply using more resources?
← Sunstone dev log 28: exposure compensation Newest